Protein generative models are often viewed as tools for designing novel sequences, but they can also be interpreted more fundamentally as data-driven maps capturing evolutionary constraints. In this perspective, sequence probabilities define an effective evolutionary landscape that links mutational effects, epistasis, and long-term protein diversification within a common framework. This talk discusses how models learned from natural sequence variation can move beyond generation and prediction to provide an interpretable description of protein evolution across scales, from single mutations in their local sequence context to the emergence of distant homologs.